NDE Data Correlation Using Encode-Decoder Networks with Scalogram Images

Abstract

Nondestructive Evaluation (NDE) technologies are increasingly used for structural condition assessments. Over the lifespan of a structure, a variety of NDE techniques may be employed, leading to a scenario where a structures life-cycle time history is depicted through a variety of complex and heterogeneous measurements. Therefore, improved understanding of the statistical associations between NDE data sources would allow engineers to integrate these data sources for analysis purposes. It would also provide new insights into the fundamental information shared between heterogeneous NDE observations, potentially leading to new forms of structural monitoring and assessment. This paper explores the correlations between NDE data types through an encoder-decoder neural network architecture. The network is designed to take in one type of NDE measurement as input, generating a synthetic measurement from a second NDE measurement as output. At the center of the encoder is a dimensionally reduced latent representation of the information that is shared between two associated NDE data sources. Additionally, this paper shows how transforming waveform NDE data into 2D time-frequency images using a Continuous Wavelet Transform (CWT) facilitates network training and representation of these shared fundamental data features. To illustrate this concept, the results from a series of laboratory scale tests are presented, representing how this network architecture would represent information collected from NDE of bridge decks.